# Full text of "An inequality for correlated measurable functions"

## See other formats

```AN INEQUALITY FOR CORRELATED MEASURABLE FUNCTIONS

FABIO ZUCCA

Abstract. A classical inequality, which is known for families of monotone functions, is generalized
to a larger class of families of measurable functions. Moreover we characterize all the families of
functions for which the equality holds. We give two applications of this result, one of them to a
problem arising from probability theory.

1. Introduction

The aim of this paper is to generahze an inequaUty, originally due to Chebyshev and then
rediscovered by Stein in [3]. Usually this result is stated for monotonic real functions: the classical
inequality is

where / and g are monotonic (in the same sense) real functions (see for instance ^ and ^ for a
more general version). If a = 6 — 1 then this inequality has a probabilistic interpretation, namely
E[/g'] — E[/]E[(/] > (where E denotes the expectation), that is, the covariance of / and g is
nonnegative.

Our approach allows us to prove the inequality for functions defined on a general measurable
space, hence we go beyond the usual ordered set R. More precisely, we prove an analogous result
for general families of measurable functions that we call correlated functions (see Definition 12.11 for
details). In particular we characterize all the families of functions for which the equality holds.

Here is the outline of the paper. In Section [5] we introduce the terminology and the main tools
needed in the sequel. In particular Sections 12. II and are devoted to the construction of an order
relation and a c-algebra on a particular quotient space. In Section [3] we state and prove our main
result (Theorem 13. 1|) which involves k correlated functions; the special case k = 2 requires weaker
assumptions (see also Remark 13. ip . We give two applications of this inequality in Section H) the
first one involves a particular class of power series, while the second one comes from probability
theory.

We start from a very general setting. Let us consider a set X, a partially ordered space (y, >y)
and a family M = {/jjier (where F is an arbitrary set) of functions in . We consider the
equivalence relation on X

and we denote by X/^ the quotient space, by [x] the equivalence class of x E X and by tt the
natural projection of X onto X/^. Roughly speaking, by means of this procedure, we identify
points in X which are not separated by the family TV.

2. Preliminaries and basic constructions

y

h{x)=fi{y), Vier

2000 Mathematics Subject Classification. 26D15, 28A25.

Key words and phrases, integral inequalities, measure, cartesian product, ordered set.

To the family M corresponds a natural counterpart A/"^ = {</)/. jjgr of functions in where,
by definition, i;^>/([x]) := /(x), for all x G X and for every / G satisfying

(2.1) Vx,yGX:x~y^/(x) = /(y)

(this holds in particular for all the functions in TV). It is clear that the family 7V^ separates the
points of X/ r^.

Given any function g defined on X/^ we denote by Hg the function g o n; observe that (pT^^ = g for
all g G and tt,^^ = / for every / satisfying equation ()2.ip . Clearly g i— > vr^ is a bijection from

y^/~ onto the subset of function in Y-^ satisfying equation (j2.ip .

Note that given /, fi G Y^ which satisfy equation ()2.ip (resp. g, g\ G Y^l-) then / >y /i
(resp. g>Y gi) implies (pf >(pfi (resp. -Kg > vr^J.

2.1. Induced order. In order to prove Theorem 13. II we cannot take advantage, as in the classical
formulation, of an order relation on the set X. Under some reasonable assumptions (se Definition l2.1l
below) we can transfer the order relation from Y to X/r^ where we already defined a family Af^
related to the original Af. This will be enough for our purposes.

Definition 2.1. The functions in M are correlated if, for all i £T and x,y € X,

(2.2) Mx) >Y fi{y) =^ fj{x) >Y /,(y), Vj G r.

We note that the definition above can be equivalently stated as follows: for all i,j G T and

X G X,

/-l((-oo,/,(x)))C/-i((-oo,/,(x)]).

Besides, if y = M with its natural order, then the functions in M are correlated if and only if for
all i,j G r and x,y G X,

(2.3) {m-m){Mx)-My))>0.

In particular if X is a totally ordered set and all the functions in J\f are nondecreasing (or nonin-
creasing) then they are correlated.

A family of correlated functions induces a natural order relation on the quotient space X/^.

Lemma 2.1. If the functions in Af are correlated then the relation on X/^

[x] >^ [y] ^ fi{x) >Y f^{y), Vi G r

is a partial order. If {Y, >y) is a totally ordered space then the same holds for (X/^, >^). Moreover
Af^ is a family of nondecreasing functions (hence they are correlated).

Proof. It is straightforward to show that >^ is a well-defined partial order (clearly it does not
depend on the choice of x (and y) within an equivalence class). We prove that, if >y is a total
order, the same holds for >^. Indeed if [x] / [y] then there exists i G T such that fi{x) 7^ fi{y);
suppose that fi{x) > fi{y) then, by equation (|2.2|) . [x] >^ [y]. It is trivial to prove that (pf- is
nondecreasing for every i G F, whence they are correlated since the space (X/^,>^) is totally
ordered. □

A subset I of an ordered set, say Y, is called an interval if and only if for all x,y & I and z gY
then X >Y z >y y implies z G I. Note that given an interval / C y then (l)J^{I) is an interval of
X/^ for every i G F.

Given x,y G X such that [x] >^ [y] we define the interval [[y],[x]) := {[z] G X/^ : [y] <
[z] < [x]}; the intervals [[y],[x]], ([y], [x]] and ([y], [x]) are defined analogously. In particular for
any x G X, we denote by [[x],+oo) and (— cxd, [x]] the intervals {[y] G X/^ : [y] >^ [x]} and
{[y] G X/^ : [x] >^ [y]} respectively.

2.2. Induced c-algebra and measure. This construction can be carried on under general as-
sumptions. Let us consider a measurable space with a positive measure {X,T,x,n) and an equiva-
lence relation ~ on X such that for all x € X and A G Sx,

(2.4) X e A^[x]CA.

There is a natural way to construct a cr-algebra on X/^, namely define

:= {7r{A) : A G Sx}

where 7r{A) '■= {[x] : x G A}. This is the largest u-algebra on X/^ such that the projection map vr
is measurable. Observe that A t^^A) is a bijection from T,x onto S^. It is natural to define a
measure fl := fi-j^ by

nHA)) = fi{A), yA G Sx.

It is well known that a function g : X/^ ^ M is measurable if and only if vr^ is measurable.
Moreover g is integrable (with respect to Jl) if and only if vr^ is integrable (with respect to /i) and

(2.5) / TTgdfj, = / gdJI.

Jx Jx/^

We say that a function g is integrable if at least one of the integrals of the two nonnegative functions
g~^ := max{g, 0) and g~ := — mm{g, 0) is finite; hence the integral of g can be unambiguously
defined as the difference of the two integrals (where =boo + z := ±00 for all 2: G M and • ±00 := 0).
This notion is sligthly weaker than the usual one: to remark the difference, when the integrals of
g~^ and g~ are both finite the function g is called summable.

It is a simple exercise to check that the equivalence relation defined in Section 12.11 satisfies
equation (j2.4p if Sx = <7(/i : ? G F) (that is, T,x is the minimal u-algebra such that all the
functions in J\f are measurable); this equivalence relation along with its induced cr-algebra and
measure will play a key role in the next section.

Remark 2.1. It is easy to show that ii h,r : X are two integrable functions such that the sum
f-^ hdfj, + f-^ rdfj, is not ambiguous (i.e. it is not true that hd^ = ±00 and f-^ rd^ = ^00) then
h-\- r is integrable and

(2.6) / {h + r)dn= / M/u + / rd/i

Jx Jx Jx

(both sides possibly being equal to ±00). This will be useful in the proof of Lemma 13.31

3. Main result

Throughout this section we consider a measurable space with finite positive measure {X, T,x , fJ-)
and a family of correlated functions M = {/ijier, where Ex = o"(/j : i G F). Let us consider
y = M with its natural order >. The equivalence relation ~, the (total) order >^ and the space
(X/^,S^,7l) are introduced according to Sections 12.11 and 12.21 It is clear that contains the
cr-algebra generated by the set of intervals {(l)J^{I) : i G F, / C M is an interval}. More precisely
it is easy to see that, by construction, all the intervals of the totally ordered set {X/^, >^) are
measurable since A/"^ separates points.

The main result is the following.

Theorem 3.1. Let ^{X) < +00.

(1) If f , g are two integrable, /i-a. e. correlated functions such that fg is integrable then

(3.1) / fgdfi > [ /d/x / gdfi.

Jx Jx Jx

Moreover, if f , g are summable, then in the previous equation the equality holds if and only
if at least one of the functions is fi-a.e constant.
(2) If {fi}i=i be a family of measurable functions on X which are nonnegative and fi-a.e. cor-
related then

(3.2) Kxf-^ n/*d/.>n/ /.d^-

Jx -^j^ -^^^ Jx

Moreover if fi^l^ ^ (0, +oo) for alii = 1, . . . ,k, then in the previous equation the equality
holds if and only if at least k — 1 functions are fi-a.e. constant.

Before proving this theorem, let us warm up with the following lemma; though it will not be
used in the proof of Theorem 13.11 nevertheless it sheds some light on the next step.

Lemma 3.2. Let M := {{a^i(i)}jeN}j=i be a family of nonnegative and nondecreasing sequences
and {/UjjjgN be i family of strictly positive real numbers. IfYlil^i < then

(3.3) y^i^i) n ^i^3)iJ-i > n ^i(.?')/^»-

i i j=l j=l i

Moreover if for every j we have < Yli^iU) < +°*-^ then the equality holds if and only if at least
k — 1 sequences are constant.

Proof. We prove the first part of the claim for two finite sequences {xj}"^^ and since the

general case follows easily by induction on k and using the Monotone Convergence Theorem as n
tends to infinity.

It is easy to prove that

n n n n

(3.4) ^/Uj^Xiyi/ij-^Xi/ii^yi/Xi = ^ {xi-Xj){yi-yj)^ifij = ^ {xi-Xj){yi-yj)^i^ij.

i=l i=l i=l i=l ij'-i^j ij'-i>j

Indeed

71 n n

^ /ii ^ Xiyifii = ^ {xiyt + Xjyj)^ij2j + ^ Xiyifi^

i=l i=l i;j'-i>j *=1

and

n n n

^ Xi^^i ^ yifii = ^ {xiyj + Xjyi)niHj + ^ Xiynxf.
This implies easily that

n n n n

^ /Xj ^ Xiyiiii - ^ Xiiii ^ yj/Xj > 0.

i=l i=l 1=1 i=l

If either at least k — 1 sequences are constant or one sequence is equal to 0, then we have an
equality. The same is true if Xj(j)/ij = +oo for some j and J2i XiU)f^i > for all j, since both
sides of equation (I3.3p are equal to +oo. On the other hand by using the first part of the theorem

and by taking the limit in equation ()3.4p as n tends to infinity, for all 1 < ii < ^2 ^

i j=l j=l i

k

>(y^^^i)^xi{il)xi{i2)^li J| ^xi{i)ixi-\Y^xi{i)iii

j¥=ji,h « i,ii:i>ii

If both {xi{ji)}i and {xi{j2)}i are nonconstant then there exist r < / and ri < li such that Xriji) <
xi{ji) and Xri{j2) < xi^{j2). This implies

in) (ii) > and Xmax(i,/i)02) -

^min(r,ri)(j2) > 0, thus the right hand side of equation (|3.5|) is strictly positive (just consider
the summation over : i > max(/,Zi),ii < min(r, ri)}) and we have a strict inequality in

equation (j3.3p . □

The proof of the previous lemma clearly suggests a second lemma which will be needed in the
proof of Theorem 13.11

Lemma 3.3. Let Af := {/, g} where f,g:X^M are two summable functions such that fg is
integrable (for instance if f and g are fi-a.e. correlated). If ^{X) < +00 then

K^) / f{x)9{x)dfi{x) = / f{x)dfx{x) / g{x)dfx{x)

(3.6)

+ 1^ / {fix) - fiy))i9ix) - 5(y))d/i(x)d/i(2/).

Proof. Note that

(3.7) fix)gix) + f{y)g{y) = fix)g{y) + f{y)g{x) + (/(x) - fiy))igix) - g{y))-

where fix)g{y) and fiy)gix) are summable on X x X, since /, 5 are summable. If we define
hix.y) := fix)g{y) + f{y)gix) and r{x,y) := (/(x) - f{y)){g{x) - g{y)) then, according to Re-
mark im we just need to prove that h and r are integrable (since /i + r is integrable by hypothesis).

If /, g are summable then, by equation (|3.7|) . fg is integrable if and only if (/(x) — f{y)){g{x) —
giy)) is integrable on X x X (since the sum of an summable function and an integrable function is
an integrable function) and equation (|3.6|) follows. Clearly if / and g are correlated then {fix) —
f{y)){g{x) — g{y)) is nonnegative thus integrable.

□

Proof of Theorem I3.il

(1) By equation (|2.5p it is enough to prove that

/ (pfCpgdjL > / (pfdjl + / (t>gdjl.

Jx/^ Jx/^ Jx/^

If / and g are summable then the claim follows from equation (j3.6p of Lemma 13.31 Oth-
erwise, without loss of generality, we may suppose that ^/d^^ = fdfi = +00. If

J^^ <j)gdj[ = gdfi < then there is nothing to prove. If f-^ gdfj, > then either g =
//-a.e. , in this case both sides of equation (13. ip are equal to 0, or there exists x £ X/^
such that 7l([x,+oo)) > and 4>f,4>g > on [x,+oo) (since and (pg are nondecreasing) .

Clearly J[^^_^_^)4)fdn = +00 and 4>f{y)4>g{y) > 4>fiy)(l>gix) for all y G [a;, +00), hence both
sides of equation (j3.ip are equal to +00.

If one of the two functions is constant then the equality holds. If / and g are nonconstant
(that is, (j)f and (pg are nonconstant) then there exists xo,yo ^ such that xq >~ yo,
(j)f{xo) > 0/(yo), 4>g{xo) > (pgiyo), 'Pi{-oo,yo]) > and 7Z([xo,+oo)) > (this can be done
as in Lemma l3.3p . Hence, using equation (j3.6p . we have that,

> ( / ihi^) - My))iM^) - My)Wi^W{y))

^ i[xo,+oo)x(-oo,yo]

>7l((-oo,2/o])7x([xo,+oo))((?:)/(2;o) - cl)f{yo)){cl)g{xo) - (t)g{yo)) > 0.
(2) Let us suppose that fi is summable for alH = 1, . . . , A;. It is enough to prove that

JX/^ i=l i=i J^l-

In the previous part of the theorem, we proved the claim for two functions and (^g\ as
in Lemma |3.'2| the general case follows by induction on k.

If at least two functions are nonconstant, say ^^^^ then as before we may find
2^0, yo e Xj^ such that xq >^ yo, 4'fi{xo) > 4'fiiyo), 4>f2{xo) > (^/aCyo), 7^((-oo,yo]) >
and 7^([xo, +00)) > (this can be done as in Lemma [3^3]) . By applying the first part of the
claim to the family (of k — 1 functions) (j)fj^(j)f^, (pj^, . . . , (f)f^ (which are clearly still correlated
since they are nondecreasing) and using equation (j3.6p we have that.

j^/^ i=i i=i j^i-^

k

= U{X/^)l (l)f^(j)f^d]I - (Phd-p- ^/ad/i) n / ^f^^'P
^ Jx/^ Jx/^ Jx/^ ' Jxj^

>

(0/i(a;) - (t>fi{y)){4>f2(.x) - '/'/2(y))dA^(a;)d^(y)j / 0/^d/i
>7l((-oo,yo])7l([xo,+oo))((/)/,(a;o) (yo))(0/2(2;o) -0/2(yo))n / ^fr^P > ^

i=3 -^^/^

since < j-^^ cpjAjiK +00 for all i = 1, ... , k, thus the second part of the claim is proved.

n

Note that if fidfi = +00 for some i and fjdfi > for all j (otherwise both sides of
equation ()3.2p are equal to 0) then both sides of equation ()3.2p are equal to +00; indeed apply
the first part of the theorem to the family of correlated bounded functions {mm{ fi,n)}^^i (where
n € N) and take the limit of both sides of equation (13. 2p as n tends to +00.

Remark 3.1. According to Theorem 13. H there is a difference between the case k = 2 and k > 2;
indeed in the latter case the inequality cannot be proved for integrable (or even summable) //-
a.e. correlated functions which are not nonnegative. Something happens in the inductive process,
namely if are correlated this may not be true for {/i/2,/3, • • • , fk} (if the functions are

not positive). Here is a counterexample: take X = [—1,1] endowed with the Lebesgue measure,
fi{x) = f2{x) ■■= xl[_ifl-\{x) and fi{x) := x - fi{x) for all i > 3.

Strictly speaking. Theorem 13 . 1 1 could be proved without the constructions of Sections l2 . 1 1 and [2^21
one has just to use carefully equation ()2.3p and Lemma [3^ Our approach simplifies the proof of
Theorem 13.11 and gives a better understanding of the role of the correlation hypothesis (compared
to the usual monotonicity) .

We finally observe that if we consider two integrable anticorrelated functions (meaning that
(/(x) — f{y)){g{x) — g{y)) < for all x,y e X) such that fg is integrable then, clearly, we have

4. Final remarks and examples

Let us apply Theorem [3] to a class of power series. We consider f{z) := Ylin=i ^nz"" where
{an}n is a sequence of nonnegative real numbers and we suppose that {p"an} is nonincreasing
(resp. nondecreasing) for some p such that < p < R (where R is the radius of convergence). Then
the function z i— > (p — z)f{z) is a nonincreasing (resp. nondecreasing) on [0,p).

Indeed if we suppose that {p^a^} is nonincreasing then, for all 2,7 such that < z < 7 < p, we
have

+00

5^a„z" = ^a„p"(z/7)"(7/p)"

n=l

ra=l

E+00 n +°°

n=l

p-z'

where, in the first inequality, we applied Theorem 13.11 to the (correlated) functions /i(n) := an/o"
and /2(n) := [zj^Y" defined on N endowed with the measure := X]nGA(7//')"- '^'^^ when
{p"a„} is nondecreasing is analogous (observe that now the functions f\ and are anticorrelated).
\i z < p < R then f\ and are nonconstant functions, hence the function 2; 1-^ (p — z)f{z) is
strictly monotone.

We draw our second application application from probability theory. To emphasize this, we
denote the measure space by {i},J-,¥) and we speak of random variables and events instead of
measurable functions and measurable sets respectively. We note that if A; = 2 then Theorem 13.11
says that correlated variables have nonnegative covariance that is, E[/i/2] — E[/i]E[/2] > (where
E[/] := /dP is the usual expectation).

We call the (real) random variables {Xq , Xi , . . . , Xf^. } independent if and only if, for every family
of Borel sets {^0, ^1, • • • , ^fc}, we have P(nf^o{Xi e ylj) = HLo ^i^i ^ where ¥{Xi € Ai) is
shorthand for F{{uj e ft : Xi{uj) e Ai}).

In order to make a specific example, let us think of the variable Xi (i = 1, . . . ,k) as the (random)
time made by the i-th contestant in an individual time trial bicycle race and let Xq be our own
(random) time; we suppose that each contestant is unaware of the results of the others (this is the
independence hypothesis). If we know the probability of winning a one-to-one race against each of
our competitors we may be interested, for instance, in estimating the probability of winning the

race. Such estimates are possible as a consequence of Theorem I3.1| indeed we have that

k

'intl{x^ > Xo}) > HnXi > Xo)

i=l
k

p(nti{Xi < Xo}) > WnXi < Xo).

i=l

Thus the events {{Xi > Xo}}f=i (resp. {{Xi < Xo}}f^ i) are positively correlated (roughly speaking
this means that knowing that {Xi > Xq} makes, for instance, the event {X2 > Xo} more likely
than before).

The proof of these inequalities is straightforward. If we define ^{A) := ¥{Xo € A) for all Borel
sets ^ C M, then, according to Fubini's Theorem,

k

>Xo)= [ F{Xi > t)dfi{t), n^tiiXi > Xo}) = [ f\nXi > t)d^(t)

JR JR ~^

k

p(x, < Xo) = / P(x, < t)dfi{t), P(nii{Xi < Xo}) = / TTP(Xi < t)dfi{t).

JR JR,,

i=l

Indeed

F{Xi > Xq) = [ di^{s)dn{t) =11 dv{s)dii{t) = [ F{Xi > t)dfi{t)

J{{s,t)eR'^:s>t} JRJ[t,+OD) JR

where i^{A) := P(Xj £ A) for all borel sets ^ C R and the first equality holds since Xj and Xo are
independent. The remaining cases are analogous. Note that {P(Xj > t)}^^^ and {P(Xj < t)}f=i are
both families of monotone (thus correlated) functions; Theorem 13.11 vields the claim. This example
can be easily extended to a more interesting case: namely when {Xi, . . . ,Xfc} have identical laws
and are independent conditioned to Xo (see Chapters 4 and 6 of [1] for details). In this case one
can prove that

k

P(n^=i{X, G A}) > ]JP(Xi eA), C E Borel set.

i=l

The proof makes use of Theorem 13.11 in its full generality but this example exceeds the purpose of
this paper.

Acknowledgments
The author thanks S. Mortola for useful discussions.

References

[1] P. Billingsley, Probability and measure, Wiley Series in Probability and Mathematical Statistics, John Wiley &
Sons, New York, 1995.

[2] R.A. Brualdi, Mathematical Notes: Comments and Complements, Amer. Math. Monthly 84, (1977), n. 10,
803-807.

[3] S.K. Stein, An inequality in two monotonic functions, Amer. Math. Monthly 83, (1976), n. 6, 469-471.

F. ZUCCA, DiPARTIMENTO DI MATEMATICA, POLITECNICO DI MiLANO, PiAZZA LEONARDO DA ViNCI 32, 20133

MiLANO, Italy.